The open-source AI model GLM 5.3 is specifically tuned for offensive capabilities like hacking